A Comparative Analysis of Deep Learning Models for Prediction of Microsatellite Instability in Colorectal Cancer

Colorectal cancer remains one of the most prevalent and fatal malignancies worldwide, underscoring the necessity for early and precise diagnostic approaches to enhance patient prognoses. This study proposes a deep learning-based model for predicting microsatellite instability (MSI) in colorectal cancer using hematoxylin and eosin (H&E)-stained histopathological tissue slides. A classification framework was constructed using convolutional neural networks (CNN) and optimized through transfer learning techniques. The dataset, comprising 150,000 unique H&E-stained histologic image patches, was sourced from an open-access Kaggle repository, with 80% allocated to training and 20% to testing. A comparative evaluation of nine pre-trained models demonstrated that the VGG19 architecture yielded the highest classification performance, achieving an accuracy of 90.60%, a precision of 88.60%, a sensitivity of 93.10%, and an AUC score of 90.60%. Considering its high performance, the proposed model is expected to assist pathologists in clinical decision-making, potentially enhancing diagnostic accuracy in real-world medical applications.